RE: Neural darwinism

From: Phil Goetz (
Date: Sun Jan 30 2005 - 23:13:16 MST

> True, but the brain is known to be highly
> inefficient in many many areas,
> e.g. (to choose a simple example) motion-direction
> detection neurons that
> are off by 80 degrees on average but are correct en
> masse due to averaging

This is not necessarily inefficiency. It may actually
be very efficient. Neurons aren't "off by 80 degrees"
so much as they have wide response zones and respond
somewhat probabilistically within those zones.
Neurons use redundancy to tolerate noise. In
exchange, they can use much, much lower voltage levels
(by about 5 orders of magnitude) than integrated
circuits do, because they don't need to be completely
reliable. This may lead to lower overall power
requirements for a computation. You could work out
the numbers for known systems, but it wouldn't be

There are also generally-accepted results showing
that noise in neural systems is beneficial, but I
don't really understand those papers, so I'll just
say that not all "noise" in the brain is bad.

> > Evaluating a
> > single prospective program requires
> > passing it through every related cortex area,
> > meaning that you need duplication not only in
> > the area being evaluated, but in ALL areas,
> > because
> > you assume all the competing programs are being
> > evaluated in parallel. Also important is that
> > you cannot evolve concepts in multiple cortical
> > areas at the same time.
> why?

I shouldn't have stated this so confidently when
it relies on my personal, not-universally-accepted
notions of how the brain works.

If you imagine that the brain processes stimuli,
and converts them into something like symbols,
and then manipulates these symbols somewhere (say
prefrontal cortex) to,
say, come up with a plan, and then pushes the
resulting plan symbols back towards the brain's
posterior (towards motor cortex) to carry out the
plan, then what I said wouldn't follow.

I do think that the brain does something like
symbol-processing, because we can productively
generate an infinite number of, for instance,
sentences fitting a language's grammar.
But I don't think these symbols are cut off
from the sense data that gives rise to them.
If this were so, non-verbal priming would not have
such dramatic effects on abstract thought. Also,
concepts would probably be more neatly defined,
along the lines of necessary and sufficient
conditions; or at least, we would have difficulty
working with non-typical category members, working
with metaphors, etc. IMHO.

I think that whenever a "symbol" is active somewhere
in anterior cortex, it is linked to, and activates (or
is activated by), a series of cortical areas
connecting it all the way back to the posterior
sensory cortex. There's some evidence for this;
primarily I'm thinking of experiments that show
that visualization activates primary visual cortex.
Damasio's work on "convergence zones" is an example
of a theory of this type. Also any theory that
involves using different frequencies to create
different "bindings" between cortical areas.

Say you're using neural darwinism in cortex area A to
select between different planned movement sequences.
If you buy into my notion of how the brain works,
that means that each competing program in area A has
to be activating all of the other cortical areas
that create/correlate with the "symbols" being used
in area A. All the way back into sensory cortex,
or at least premotor cortex. And to prevent
collisions, that means that EVERY SINGLE CONCEPT
stored in every cortical area needs to be stored
hundreds of times over, so that these hundreds of
copies can be activated separately and used
simultaneously by competing programs in your planning

(I'm not talking about the ordinary sort of redundancy
in the brain. Each of these hundreds of copies has
got to have that level of redundancy.)

You could probably calculate whether this is possible.
We have some "maps" of temporal cortex indicating
"words" of various types are stored. You could
probably estimate the number of different "words"
stored in a particular area of cortex, and estimate
the number of neurons in that area, and figure out
whether or not the number of neurons is in fact at
least a hundred times as many as you would need to
store those words. But, again, this calculation
would not be trivial.

I should probably re-read this before sending,
but I'm too tired.

- Phil G.

Do you Yahoo!?
Yahoo! Mail - Find what you need with new enhanced search.

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:50 MDT